Normative Profiles in Reinforcement Learning

Autonomous agents in open contexts face normative dilemmas not only in the presence of explicit norm conflicts, but every time they balance normative requirements with task completion. This discourages attempts to model artificial agents as implicit moral agents, who always follow norms. In this paper, we introduce the concept of normative profiles and investigate their role in ex ante compliance choices. We take the perspective of a single agent facing the choice of whether to comply or not with a norm. We then propose a way to emulate an intermediate behaviour between a pure utilitarian and a pure deontologist profile within a reinforcement learning architecture. We assess the necessity of having two specific parameters, the reluctance and the prudence levels, through four experimental scenarios covering the ensemble of the normative configurations. Finally, we show that, by altering these parameters, it is possible to make the agent shift its behaviour from utilitarian to deontologist, and vice versa.